1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 *  linux/mm/swap_state.c
  4 *
  5 *  Copyright (C) 1991, 1992, 1993, 1994  Linus Torvalds
  6 *  Swap reorganised 29.12.95, Stephen Tweedie
  7 *
  8 *  Rewritten to use page cache, (C) 1998 Stephen Tweedie
  9 */
 
 10#include <linux/mm.h>
 11#include <linux/gfp.h>
 12#include <linux/kernel_stat.h>
 13#include <linux/swap.h>
 14#include <linux/swapops.h>
 15#include <linux/init.h>
 16#include <linux/pagemap.h>
 
 17#include <linux/backing-dev.h>
 18#include <linux/blkdev.h>
 19#include <linux/pagevec.h>
 20#include <linux/migrate.h>
 21#include <linux/vmalloc.h>
 22#include <linux/swap_slots.h>
 23#include <linux/huge_mm.h>
 24
 25#include <asm/pgtable.h>
 26
 27/*
 28 * swapper_space is a fiction, retained to simplify the path through
 29 * vmscan's shrink_page_list.
 30 */
 31static const struct address_space_operations swap_aops = {
 32	.writepage	= swap_writepage,
 33	.set_page_dirty	= swap_set_page_dirty,
 34#ifdef CONFIG_MIGRATION
 35	.migratepage	= migrate_page,
 36#endif
 37};
 38
 39struct address_space *swapper_spaces[MAX_SWAPFILES] __read_mostly;
 40static unsigned int nr_swapper_spaces[MAX_SWAPFILES] __read_mostly;
 41static bool enable_vma_readahead __read_mostly = true;
 42
 43#define SWAP_RA_WIN_SHIFT	(PAGE_SHIFT / 2)
 44#define SWAP_RA_HITS_MASK	((1UL << SWAP_RA_WIN_SHIFT) - 1)
 45#define SWAP_RA_HITS_MAX	SWAP_RA_HITS_MASK
 46#define SWAP_RA_WIN_MASK	(~PAGE_MASK & ~SWAP_RA_HITS_MASK)
 47
 48#define SWAP_RA_HITS(v)		((v) & SWAP_RA_HITS_MASK)
 49#define SWAP_RA_WIN(v)		(((v) & SWAP_RA_WIN_MASK) >> SWAP_RA_WIN_SHIFT)
 50#define SWAP_RA_ADDR(v)		((v) & PAGE_MASK)
 51
 52#define SWAP_RA_VAL(addr, win, hits)				\
 53	(((addr) & PAGE_MASK) |					\
 54	 (((win) << SWAP_RA_WIN_SHIFT) & SWAP_RA_WIN_MASK) |	\
 55	 ((hits) & SWAP_RA_HITS_MASK))
 56
 57/* Initial readahead hits is 4 to start up with a small window */
 58#define GET_SWAP_RA_VAL(vma)					\
 59	(atomic_long_read(&(vma)->swap_readahead_info) ? : 4)
 60
 61#define INC_CACHE_INFO(x)	do { swap_cache_info.x++; } while (0)
 62#define ADD_CACHE_INFO(x, nr)	do { swap_cache_info.x += (nr); } while (0)
 63
 64static struct {
 65	unsigned long add_total;
 66	unsigned long del_total;
 67	unsigned long find_success;
 68	unsigned long find_total;
 69} swap_cache_info;
 70
 71unsigned long total_swapcache_pages(void)
 72{
 73	unsigned int i, j, nr;
 74	unsigned long ret = 0;
 75	struct address_space *spaces;
 76
 77	rcu_read_lock();
 78	for (i = 0; i < MAX_SWAPFILES; i++) {
 79		/*
 80		 * The corresponding entries in nr_swapper_spaces and
 81		 * swapper_spaces will be reused only after at least
 82		 * one grace period.  So it is impossible for them
 83		 * belongs to different usage.
 84		 */
 85		nr = nr_swapper_spaces[i];
 86		spaces = rcu_dereference(swapper_spaces[i]);
 87		if (!nr || !spaces)
 88			continue;
 89		for (j = 0; j < nr; j++)
 90			ret += spaces[j].nrpages;
 91	}
 92	rcu_read_unlock();
 93	return ret;
 94}
 95
 96static atomic_t swapin_readahead_hits = ATOMIC_INIT(4);
 97
 98void show_swap_cache_info(void)
 99{
100	printk("%lu pages in swap cache\n", total_swapcache_pages());
101	printk("Swap cache stats: add %lu, delete %lu, find %lu/%lu\n",
102		swap_cache_info.add_total, swap_cache_info.del_total,
103		swap_cache_info.find_success, swap_cache_info.find_total);
104	printk("Free swap  = %ldkB\n",
105		get_nr_swap_pages() << (PAGE_SHIFT - 10));
106	printk("Total swap = %lukB\n", total_swap_pages << (PAGE_SHIFT - 10));
107}
108
109/*
110 * __add_to_swap_cache resembles add_to_page_cache_locked on swapper_space,
111 * but sets SwapCache flag and private instead of mapping and index.
112 */
113int __add_to_swap_cache(struct page *page, swp_entry_t entry)
114{
115	int error, i, nr = hpage_nr_pages(page);
116	struct address_space *address_space;
117	pgoff_t idx = swp_offset(entry);
118
119	VM_BUG_ON_PAGE(!PageLocked(page), page);
120	VM_BUG_ON_PAGE(PageSwapCache(page), page);
121	VM_BUG_ON_PAGE(!PageSwapBacked(page), page);
122
123	page_ref_add(page, nr);
124	SetPageSwapCache(page);
 
125
126	address_space = swap_address_space(entry);
127	xa_lock_irq(&address_space->i_pages);
128	for (i = 0; i < nr; i++) {
129		set_page_private(page + i, entry.val + i);
130		error = radix_tree_insert(&address_space->i_pages,
131					  idx + i, page + i);
132		if (unlikely(error))
133			break;
134	}
135	if (likely(!error)) {
136		address_space->nrpages += nr;
137		__mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, nr);
138		ADD_CACHE_INFO(add_total, nr);
139	} else {
 
 
 
140		/*
141		 * Only the context which have set SWAP_HAS_CACHE flag
142		 * would call add_to_swap_cache().
143		 * So add_to_swap_cache() doesn't returns -EEXIST.
144		 */
145		VM_BUG_ON(error == -EEXIST);
146		set_page_private(page + i, 0UL);
147		while (i--) {
148			radix_tree_delete(&address_space->i_pages, idx + i);
149			set_page_private(page + i, 0UL);
150		}
151		ClearPageSwapCache(page);
152		page_ref_sub(page, nr);
153	}
154	xa_unlock_irq(&address_space->i_pages);
155
156	return error;
157}
158
159
160int add_to_swap_cache(struct page *page, swp_entry_t entry, gfp_t gfp_mask)
161{
162	int error;
163
164	error = radix_tree_maybe_preload_order(gfp_mask, compound_order(page));
165	if (!error) {
166		error = __add_to_swap_cache(page, entry);
167		radix_tree_preload_end();
168	}
169	return error;
170}
171
172/*
173 * This must be called only on pages that have
174 * been verified to be in the swap cache.
175 */
176void __delete_from_swap_cache(struct page *page)
177{
178	struct address_space *address_space;
179	int i, nr = hpage_nr_pages(page);
180	swp_entry_t entry;
181	pgoff_t idx;
182
183	VM_BUG_ON_PAGE(!PageLocked(page), page);
184	VM_BUG_ON_PAGE(!PageSwapCache(page), page);
185	VM_BUG_ON_PAGE(PageWriteback(page), page);
186
187	entry.val = page_private(page);
188	address_space = swap_address_space(entry);
189	idx = swp_offset(entry);
190	for (i = 0; i < nr; i++) {
191		radix_tree_delete(&address_space->i_pages, idx + i);
192		set_page_private(page + i, 0);
193	}
194	ClearPageSwapCache(page);
195	address_space->nrpages -= nr;
196	__mod_node_page_state(page_pgdat(page), NR_FILE_PAGES, -nr);
197	ADD_CACHE_INFO(del_total, nr);
198}
199
200/**
201 * add_to_swap - allocate swap space for a page
202 * @page: page we want to move to swap
203 *
204 * Allocate swap space for the page and add the page to the
205 * swap cache.  Caller needs to hold the page lock. 
206 */
207int add_to_swap(struct page *page)
208{
209	swp_entry_t entry;
210	int err;
211
212	VM_BUG_ON_PAGE(!PageLocked(page), page);
213	VM_BUG_ON_PAGE(!PageUptodate(page), page);
214
215	entry = get_swap_page(page);
216	if (!entry.val)
217		return 0;
218
219	if (mem_cgroup_try_charge_swap(page, entry))
220		goto fail;
 
 
 
221
222	/*
223	 * Radix-tree node allocations from PF_MEMALLOC contexts could
224	 * completely exhaust the page allocator. __GFP_NOMEMALLOC
225	 * stops emergency reserves from being allocated.
226	 *
227	 * TODO: this could cause a theoretical memory reclaim
228	 * deadlock in the swap out path.
229	 */
230	/*
231	 * Add it to the swap cache.
232	 */
233	err = add_to_swap_cache(page, entry,
234			__GFP_HIGH|__GFP_NOMEMALLOC|__GFP_NOWARN);
235	/* -ENOMEM radix-tree allocation failure */
236	if (err)
 
 
 
237		/*
238		 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
239		 * clear SWAP_HAS_CACHE flag.
240		 */
241		goto fail;
242	/*
243	 * Normally the page will be dirtied in unmap because its pte should be
244	 * dirty. A special case is MADV_FREE page. The page'e pte could have
245	 * dirty bit cleared but the page's SwapBacked bit is still set because
246	 * clearing the dirty bit and SwapBacked bit has no lock protected. For
247	 * such page, unmap will not set dirty bit for it, so page reclaim will
248	 * not write the page out. This can cause data corruption when the page
249	 * is swap in later. Always setting the dirty bit for the page solves
250	 * the problem.
251	 */
252	set_page_dirty(page);
253
254	return 1;
255
256fail:
257	put_swap_page(page, entry);
258	return 0;
259}
260
261/*
262 * This must be called only on pages that have
263 * been verified to be in the swap cache and locked.
264 * It will never put the page into the free list,
265 * the caller has a reference on the page.
266 */
267void delete_from_swap_cache(struct page *page)
268{
269	swp_entry_t entry;
270	struct address_space *address_space;
271
272	entry.val = page_private(page);
273
274	address_space = swap_address_space(entry);
275	xa_lock_irq(&address_space->i_pages);
276	__delete_from_swap_cache(page);
277	xa_unlock_irq(&address_space->i_pages);
278
279	put_swap_page(page, entry);
280	page_ref_sub(page, hpage_nr_pages(page));
281}
282
283/* 
284 * If we are the only user, then try to free up the swap cache. 
285 * 
286 * Its ok to check for PageSwapCache without the page lock
287 * here because we are going to recheck again inside
288 * try_to_free_swap() _with_ the lock.
289 * 					- Marcelo
290 */
291static inline void free_swap_cache(struct page *page)
292{
293	if (PageSwapCache(page) && !page_mapped(page) && trylock_page(page)) {
294		try_to_free_swap(page);
295		unlock_page(page);
296	}
297}
298
299/* 
300 * Perform a free_page(), also freeing any swap cache associated with
301 * this page if it is the last user of the page.
302 */
303void free_page_and_swap_cache(struct page *page)
304{
305	free_swap_cache(page);
306	if (!is_huge_zero_page(page))
307		put_page(page);
308}
309
310/*
311 * Passed an array of pages, drop them all from swapcache and then release
312 * them.  They are removed from the LRU and freed if this is their last use.
313 */
314void free_pages_and_swap_cache(struct page **pages, int nr)
315{
316	struct page **pagep = pages;
317	int i;
318
319	lru_add_drain();
320	for (i = 0; i < nr; i++)
321		free_swap_cache(pagep[i]);
322	release_pages(pagep, nr);
323}
324
325static inline bool swap_use_vma_readahead(void)
326{
327	return READ_ONCE(enable_vma_readahead) && !atomic_read(&nr_rotate_swap);
 
 
328}
329
330/*
331 * Lookup a swap entry in the swap cache. A found page will be returned
332 * unlocked and with its refcount incremented - we rely on the kernel
333 * lock getting page table operations atomic even if we drop the page
334 * lock before returning.
335 */
336struct page *lookup_swap_cache(swp_entry_t entry, struct vm_area_struct *vma,
337			       unsigned long addr)
338{
339	struct page *page;
340
341	page = find_get_page(swap_address_space(entry), swp_offset(entry));
342
343	INC_CACHE_INFO(find_total);
344	if (page) {
345		bool vma_ra = swap_use_vma_readahead();
346		bool readahead;
347
 
348		INC_CACHE_INFO(find_success);
349		/*
350		 * At the moment, we don't support PG_readahead for anon THP
351		 * so let's bail out rather than confusing the readahead stat.
352		 */
353		if (unlikely(PageTransCompound(page)))
354			return page;
355
356		readahead = TestClearPageReadahead(page);
357		if (vma && vma_ra) {
358			unsigned long ra_val;
359			int win, hits;
360
361			ra_val = GET_SWAP_RA_VAL(vma);
362			win = SWAP_RA_WIN(ra_val);
363			hits = SWAP_RA_HITS(ra_val);
364			if (readahead)
365				hits = min_t(int, hits + 1, SWAP_RA_HITS_MAX);
366			atomic_long_set(&vma->swap_readahead_info,
367					SWAP_RA_VAL(addr, win, hits));
368		}
369
370		if (readahead) {
371			count_vm_event(SWAP_RA_HIT);
372			if (!vma || !vma_ra)
373				atomic_inc(&swapin_readahead_hits);
374		}
375	}
376
 
377	return page;
378}
379
380struct page *__read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
381			struct vm_area_struct *vma, unsigned long addr,
382			bool *new_page_allocated)
 
 
 
 
 
383{
384	struct page *found_page, *new_page = NULL;
385	struct address_space *swapper_space = swap_address_space(entry);
386	int err;
387	*new_page_allocated = false;
388
389	do {
390		/*
391		 * First check the swap cache.  Since this is normally
392		 * called after lookup_swap_cache() failed, re-calling
393		 * that would confuse statistics.
394		 */
395		found_page = find_get_page(swapper_space, swp_offset(entry));
396		if (found_page)
397			break;
398
399		/*
400		 * Just skip read ahead for unused swap slot.
401		 * During swap_off when swap_slot_cache is disabled,
402		 * we have to handle the race between putting
403		 * swap entry in swap cache and marking swap slot
404		 * as SWAP_HAS_CACHE.  That's done in later part of code or
405		 * else swap_off will be aborted if we return NULL.
406		 */
407		if (!__swp_swapcount(entry) && swap_slot_cache_enabled)
408			break;
409
410		/*
411		 * Get a new page to read into from swap.
412		 */
413		if (!new_page) {
414			new_page = alloc_page_vma(gfp_mask, vma, addr);
415			if (!new_page)
416				break;		/* Out of memory */
417		}
418
419		/*
420		 * call radix_tree_preload() while we can wait.
421		 */
422		err = radix_tree_maybe_preload(gfp_mask & GFP_KERNEL);
423		if (err)
424			break;
425
426		/*
427		 * Swap entry may have been freed since our caller observed it.
428		 */
429		err = swapcache_prepare(entry);
430		if (err == -EEXIST) {
431			radix_tree_preload_end();
432			/*
433			 * We might race against get_swap_page() and stumble
434			 * across a SWAP_HAS_CACHE swap_map entry whose page
435			 * has not been brought into the swapcache yet.
436			 */
437			cond_resched();
438			continue;
439		}
440		if (err) {		/* swp entry is obsolete ? */
441			radix_tree_preload_end();
442			break;
443		}
444
445		/* May fail (-ENOMEM) if radix-tree node allocation failed. */
446		__SetPageLocked(new_page);
447		__SetPageSwapBacked(new_page);
448		err = __add_to_swap_cache(new_page, entry);
449		if (likely(!err)) {
450			radix_tree_preload_end();
451			/*
452			 * Initiate read into locked page and return.
453			 */
454			lru_cache_add_anon(new_page);
455			*new_page_allocated = true;
456			return new_page;
457		}
458		radix_tree_preload_end();
459		__ClearPageLocked(new_page);
 
460		/*
461		 * add_to_swap_cache() doesn't return -EEXIST, so we can safely
462		 * clear SWAP_HAS_CACHE flag.
463		 */
464		put_swap_page(new_page, entry);
465	} while (err != -ENOMEM);
466
467	if (new_page)
468		put_page(new_page);
469	return found_page;
470}
471
472/*
473 * Locate a page of swap in physical memory, reserving swap cache space
474 * and reading the disk if it is not already cached.
475 * A failure return means that either the page allocation failed or that
476 * the swap entry is no longer in use.
477 */
478struct page *read_swap_cache_async(swp_entry_t entry, gfp_t gfp_mask,
479		struct vm_area_struct *vma, unsigned long addr, bool do_poll)
480{
481	bool page_was_allocated;
482	struct page *retpage = __read_swap_cache_async(entry, gfp_mask,
483			vma, addr, &page_was_allocated);
484
485	if (page_was_allocated)
486		swap_readpage(retpage, do_poll);
487
488	return retpage;
489}
490
491static unsigned int __swapin_nr_pages(unsigned long prev_offset,
492				      unsigned long offset,
493				      int hits,
494				      int max_pages,
495				      int prev_win)
496{
497	unsigned int pages, last_ra;
498
499	/*
500	 * This heuristic has been found to work well on both sequential and
501	 * random loads, swapping to hard disk or to SSD: please don't ask
502	 * what the "+ 2" means, it just happens to work well, that's all.
503	 */
504	pages = hits + 2;
505	if (pages == 2) {
506		/*
507		 * We can have no readahead hits to judge by: but must not get
508		 * stuck here forever, so check for an adjacent offset instead
509		 * (and don't even bother to check whether swap type is same).
510		 */
511		if (offset != prev_offset + 1 && offset != prev_offset - 1)
512			pages = 1;
513	} else {
514		unsigned int roundup = 4;
515		while (roundup < pages)
516			roundup <<= 1;
517		pages = roundup;
518	}
519
520	if (pages > max_pages)
521		pages = max_pages;
522
523	/* Don't shrink readahead too fast */
524	last_ra = prev_win / 2;
525	if (pages < last_ra)
526		pages = last_ra;
527
528	return pages;
529}
530
531static unsigned long swapin_nr_pages(unsigned long offset)
532{
533	static unsigned long prev_offset;
534	unsigned int hits, pages, max_pages;
535	static atomic_t last_readahead_pages;
536
537	max_pages = 1 << READ_ONCE(page_cluster);
538	if (max_pages <= 1)
539		return 1;
540
541	hits = atomic_xchg(&swapin_readahead_hits, 0);
542	pages = __swapin_nr_pages(prev_offset, offset, hits, max_pages,
543				  atomic_read(&last_readahead_pages));
544	if (!hits)
545		prev_offset = offset;
546	atomic_set(&last_readahead_pages, pages);
547
548	return pages;
549}
550
551/**
552 * swap_cluster_readahead - swap in pages in hope we need them soon
553 * @entry: swap entry of this memory
554 * @gfp_mask: memory allocation flags
555 * @vmf: fault information
 
556 *
557 * Returns the struct page for entry and addr, after queueing swapin.
558 *
559 * Primitive swap readahead code. We simply read an aligned block of
560 * (1 << page_cluster) entries in the swap area. This method is chosen
561 * because it doesn't cost us any seek time.  We also make sure to queue
562 * the 'original' request together with the readahead ones...
563 *
564 * This has been extended to use the NUMA policies from the mm triggering
565 * the readahead.
566 *
567 * Caller must hold down_read on the vma->vm_mm if vmf->vma is not NULL.
568 */
569struct page *swap_cluster_readahead(swp_entry_t entry, gfp_t gfp_mask,
570				struct vm_fault *vmf)
571{
 
572	struct page *page;
573	unsigned long entry_offset = swp_offset(entry);
574	unsigned long offset = entry_offset;
575	unsigned long start_offset, end_offset;
576	unsigned long mask;
577	struct swap_info_struct *si = swp_swap_info(entry);
578	struct blk_plug plug;
579	bool do_poll = true, page_allocated;
580	struct vm_area_struct *vma = vmf->vma;
581	unsigned long addr = vmf->address;
582
583	mask = swapin_nr_pages(offset) - 1;
584	if (!mask)
585		goto skip;
586
587	do_poll = false;
588	/* Read a page_cluster sized and aligned cluster around offset. */
589	start_offset = offset & ~mask;
590	end_offset = offset | mask;
591	if (!start_offset)	/* First page is swap header. */
592		start_offset++;
593	if (end_offset >= si->max)
594		end_offset = si->max - 1;
595
596	blk_start_plug(&plug);
597	for (offset = start_offset; offset <= end_offset ; offset++) {
 
 
 
 
 
 
 
598		/* Ok, do the async read-ahead now */
599		page = __read_swap_cache_async(
600			swp_entry(swp_type(entry), offset),
601			gfp_mask, vma, addr, &page_allocated);
602		if (!page)
603			continue;
604		if (page_allocated) {
605			swap_readpage(page, false);
606			if (offset != entry_offset) {
607				SetPageReadahead(page);
608				count_vm_event(SWAP_RA);
609			}
610		}
611		put_page(page);
612	}
613	blk_finish_plug(&plug);
614
615	lru_add_drain();	/* Push any new pages onto the LRU now */
616skip:
617	return read_swap_cache_async(entry, gfp_mask, vma, addr, do_poll);
618}
619
620int init_swap_address_space(unsigned int type, unsigned long nr_pages)
621{
622	struct address_space *spaces, *space;
623	unsigned int i, nr;
624
625	nr = DIV_ROUND_UP(nr_pages, SWAP_ADDRESS_SPACE_PAGES);
626	spaces = kvzalloc(sizeof(struct address_space) * nr, GFP_KERNEL);
627	if (!spaces)
628		return -ENOMEM;
629	for (i = 0; i < nr; i++) {
630		space = spaces + i;
631		INIT_RADIX_TREE(&space->i_pages, GFP_ATOMIC|__GFP_NOWARN);
632		atomic_set(&space->i_mmap_writable, 0);
633		space->a_ops = &swap_aops;
634		/* swap cache doesn't use writeback related tags */
635		mapping_set_no_writeback_tags(space);
636	}
637	nr_swapper_spaces[type] = nr;
638	rcu_assign_pointer(swapper_spaces[type], spaces);
639
640	return 0;
641}
642
643void exit_swap_address_space(unsigned int type)
644{
645	struct address_space *spaces;
646
647	spaces = swapper_spaces[type];
648	nr_swapper_spaces[type] = 0;
649	rcu_assign_pointer(swapper_spaces[type], NULL);
650	synchronize_rcu();
651	kvfree(spaces);
652}
653
654static inline void swap_ra_clamp_pfn(struct vm_area_struct *vma,
655				     unsigned long faddr,
656				     unsigned long lpfn,
657				     unsigned long rpfn,
658				     unsigned long *start,
659				     unsigned long *end)
660{
661	*start = max3(lpfn, PFN_DOWN(vma->vm_start),
662		      PFN_DOWN(faddr & PMD_MASK));
663	*end = min3(rpfn, PFN_DOWN(vma->vm_end),
664		    PFN_DOWN((faddr & PMD_MASK) + PMD_SIZE));
665}
666
667static void swap_ra_info(struct vm_fault *vmf,
668			struct vma_swap_readahead *ra_info)
669{
670	struct vm_area_struct *vma = vmf->vma;
671	unsigned long ra_val;
672	swp_entry_t entry;
673	unsigned long faddr, pfn, fpfn;
674	unsigned long start, end;
675	pte_t *pte, *orig_pte;
676	unsigned int max_win, hits, prev_win, win, left;
677#ifndef CONFIG_64BIT
678	pte_t *tpte;
679#endif
680
681	max_win = 1 << min_t(unsigned int, READ_ONCE(page_cluster),
682			     SWAP_RA_ORDER_CEILING);
683	if (max_win == 1) {
684		ra_info->win = 1;
685		return;
686	}
687
688	faddr = vmf->address;
689	orig_pte = pte = pte_offset_map(vmf->pmd, faddr);
690	entry = pte_to_swp_entry(*pte);
691	if ((unlikely(non_swap_entry(entry)))) {
692		pte_unmap(orig_pte);
693		return;
694	}
695
696	fpfn = PFN_DOWN(faddr);
697	ra_val = GET_SWAP_RA_VAL(vma);
698	pfn = PFN_DOWN(SWAP_RA_ADDR(ra_val));
699	prev_win = SWAP_RA_WIN(ra_val);
700	hits = SWAP_RA_HITS(ra_val);
701	ra_info->win = win = __swapin_nr_pages(pfn, fpfn, hits,
702					       max_win, prev_win);
703	atomic_long_set(&vma->swap_readahead_info,
704			SWAP_RA_VAL(faddr, win, 0));
705
706	if (win == 1) {
707		pte_unmap(orig_pte);
708		return;
709	}
710
711	/* Copy the PTEs because the page table may be unmapped */
712	if (fpfn == pfn + 1)
713		swap_ra_clamp_pfn(vma, faddr, fpfn, fpfn + win, &start, &end);
714	else if (pfn == fpfn + 1)
715		swap_ra_clamp_pfn(vma, faddr, fpfn - win + 1, fpfn + 1,
716				  &start, &end);
717	else {
718		left = (win - 1) / 2;
719		swap_ra_clamp_pfn(vma, faddr, fpfn - left, fpfn + win - left,
720				  &start, &end);
721	}
722	ra_info->nr_pte = end - start;
723	ra_info->offset = fpfn - start;
724	pte -= ra_info->offset;
725#ifdef CONFIG_64BIT
726	ra_info->ptes = pte;
727#else
728	tpte = ra_info->ptes;
729	for (pfn = start; pfn != end; pfn++)
730		*tpte++ = *pte++;
731#endif
732	pte_unmap(orig_pte);
733}
734
735static struct page *swap_vma_readahead(swp_entry_t fentry, gfp_t gfp_mask,
736				       struct vm_fault *vmf)
737{
738	struct blk_plug plug;
739	struct vm_area_struct *vma = vmf->vma;
740	struct page *page;
741	pte_t *pte, pentry;
742	swp_entry_t entry;
743	unsigned int i;
744	bool page_allocated;
745	struct vma_swap_readahead ra_info = {0,};
746
747	swap_ra_info(vmf, &ra_info);
748	if (ra_info.win == 1)
749		goto skip;
750
751	blk_start_plug(&plug);
752	for (i = 0, pte = ra_info.ptes; i < ra_info.nr_pte;
753	     i++, pte++) {
754		pentry = *pte;
755		if (pte_none(pentry))
756			continue;
757		if (pte_present(pentry))
758			continue;
759		entry = pte_to_swp_entry(pentry);
760		if (unlikely(non_swap_entry(entry)))
761			continue;
762		page = __read_swap_cache_async(entry, gfp_mask, vma,
763					       vmf->address, &page_allocated);
764		if (!page)
765			continue;
766		if (page_allocated) {
767			swap_readpage(page, false);
768			if (i != ra_info.offset) {
769				SetPageReadahead(page);
770				count_vm_event(SWAP_RA);
771			}
772		}
773		put_page(page);
774	}
775	blk_finish_plug(&plug);
776	lru_add_drain();
777skip:
778	return read_swap_cache_async(fentry, gfp_mask, vma, vmf->address,
779				     ra_info.win == 1);
780}
781
782/**
783 * swapin_readahead - swap in pages in hope we need them soon
784 * @entry: swap entry of this memory
785 * @gfp_mask: memory allocation flags
786 * @vmf: fault information
787 *
788 * Returns the struct page for entry and addr, after queueing swapin.
789 *
790 * It's a main entry function for swap readahead. By the configuration,
791 * it will read ahead blocks by cluster-based(ie, physical disk based)
792 * or vma-based(ie, virtual address based on faulty address) readahead.
793 */
794struct page *swapin_readahead(swp_entry_t entry, gfp_t gfp_mask,
795				struct vm_fault *vmf)
796{
797	return swap_use_vma_readahead() ?
798			swap_vma_readahead(entry, gfp_mask, vmf) :
799			swap_cluster_readahead(entry, gfp_mask, vmf);
800}
801
802#ifdef CONFIG_SYSFS
803static ssize_t vma_ra_enabled_show(struct kobject *kobj,
804				     struct kobj_attribute *attr, char *buf)
805{
806	return sprintf(buf, "%s\n", enable_vma_readahead ? "true" : "false");
807}
808static ssize_t vma_ra_enabled_store(struct kobject *kobj,
809				      struct kobj_attribute *attr,
810				      const char *buf, size_t count)
811{
812	if (!strncmp(buf, "true", 4) || !strncmp(buf, "1", 1))
813		enable_vma_readahead = true;
814	else if (!strncmp(buf, "false", 5) || !strncmp(buf, "0", 1))
815		enable_vma_readahead = false;
816	else
817		return -EINVAL;
818
819	return count;
820}
821static struct kobj_attribute vma_ra_enabled_attr =
822	__ATTR(vma_ra_enabled, 0644, vma_ra_enabled_show,
823	       vma_ra_enabled_store);
824
825static struct attribute *swap_attrs[] = {
826	&vma_ra_enabled_attr.attr,
827	NULL,
828};
829
830static struct attribute_group swap_attr_group = {
831	.attrs = swap_attrs,
832};
833
834static int __init swap_init_sysfs(void)
835{
836	int err;
837	struct kobject *swap_kobj;
838
839	swap_kobj = kobject_create_and_add("swap", mm_kobj);
840	if (!swap_kobj) {
841		pr_err("failed to create swap kobject\n");
842		return -ENOMEM;
843	}
844	err = sysfs_create_group(swap_kobj, &swap_attr_group);
845	if (err) {
846		pr_err("failed to register swap group\n");
847		goto delete_obj;
848	}
849	return 0;
850
851delete_obj:
852	kobject_put(swap_kobj);
853	return err;
854}
855subsys_initcall(swap_init_sysfs);
856#endif